Re: Q: smp.c && barriers (Was: [PATCH 1/4] generic-smp: remove single ipi fallback for smp_call_function_many())

[Oleg Nesterov <oleg@redhat.com>]

On 02/16, Peter Zijlstra wrote:
>
> On Mon, 2009-02-16 at 21:49 +0100, Oleg Nesterov wrote:
> > I am trying to understand the barriers in smp.c, please help!
> >
> > "generic-ipi: fix the smp_mb() placement" commit
> > 561920a0d2bb6d63343e83acfd784c0a77bd28d1 added smp_read_barrier_depends()
> > to generic_smp_call_function_single_interrupt().
> >
> > Why it is needed? The comment says:
> >
> > 	/*
> > 	 * Need to see other stores to list head for checking whether
> > 	 * list is empty without holding q->lock
> > 	 */
> > 	smp_read_barrier_depends();
> > 	while (!list_empty(&q->list)) {
> >
> > But we can't miss the addition to the call_single_queue.list,
> > if generic_exec_single() sees list_empty(&dst->list) it sends
> > another IPI?
>
> I was about to write a response, but found it to be a justification for
> the read_barrier_depends() at the end of the loop.

I forgot to mention I don't understand the read_barrier_depends() at the
end of the loop as well ;)

> > This commit also removed the barrier from csd_flag_wait(), is this OK?
> > Without the barrier, csd_flag_wait() can return before we see the result
> > of data->func() ?
> >
> > IOW,
> > 	int VAR = 0;
> >
> > 	void func(coid *unused)
> > 	{
> > 		VAR = 1;
> > 	}
> >
> > Now,
> >
> > 	smp_call_function_single(0, func, NULL, 1);
> > 	BUG_ON(VAR == 0);
> >
> > afaics, the BUG_ON() above is possible. Is this OK ?
>
> Would it not be the caller's responsibility to provide the needed
> serialization in this case?

Yes, yes, I don't claim this is necessary wrong. I am just asking,
is the lack of "implicit" serialization is "by design" or this was
an oversight.

Oleg.